HOMOMORPHISMS OF MULTIPLICATIVE GROUPS OF FIELDS PRESERVING ALGEBRAIC DEPENDENCE

Similar documents
COMMUTING ELEMENTS IN GALOIS GROUPS OF FUNCTION FIELDS. Fedor Bogomolov and Yuri Tschinkel

RECONSTRUCTION OF FUNCTION FIELDS. Fedor Bogomolov and Yuri Tschinkel

RECONSTRUCTION OF HIGHER-DIMENSIONAL FUNCTION FIELDS

Course 311: Michaelmas Term 2005 Part III: Topics in Commutative Algebra

MATH 8253 ALGEBRAIC GEOMETRY WEEK 12

Notes on p-divisible Groups

Groups of Prime Power Order with Derived Subgroup of Prime Order

Algebraic function fields

Rings and groups. Ya. Sysak

On Mordell-Lang in Algebraic Groups of Unipotent Rank 1

Math 145. Codimension

A Version of the Grothendieck Conjecture for p-adic Local Fields

Math 121 Homework 5: Notes on Selected Problems

Pacific Journal of Mathematics

This is a closed subset of X Y, by Proposition 6.5(b), since it is equal to the inverse image of the diagonal under the regular map:

MATH32062 Notes. 1 Affine algebraic varieties. 1.1 Definition of affine algebraic varieties

Finite affine planes in projective spaces

1 Notations and Statement of the Main Results

SPACES OF RATIONAL CURVES IN COMPLETE INTERSECTIONS

Notes on Ordered Sets

10. Smooth Varieties. 82 Andreas Gathmann

Math 203A - Solution Set 3

ZEROS OF SPARSE POLYNOMIALS OVER LOCAL FIELDS OF CHARACTERISTIC p

CHEVALLEY S THEOREM AND COMPLETE VARIETIES

5 Dedekind extensions

Reid 5.2. Describe the irreducible components of V (J) for J = (y 2 x 4, x 2 2x 3 x 2 y + 2xy + y 2 y) in k[x, y, z]. Here k is algebraically closed.

Lecture 7: Etale Fundamental Group - Examples

A finite universal SAGBI basis for the kernel of a derivation. Osaka Journal of Mathematics. 41(4) P.759-P.792

ALGEBRAIC GEOMETRY (NMAG401) Contents. 2. Polynomial and rational maps 9 3. Hilbert s Nullstellensatz and consequences 23 References 30

SPACES OF RATIONAL CURVES ON COMPLETE INTERSECTIONS

Math 210B. Artin Rees and completions

5 Dedekind extensions

4. Images of Varieties Given a morphism f : X Y of quasi-projective varieties, a basic question might be to ask what is the image of a closed subset

Math 429/581 (Advanced) Group Theory. Summary of Definitions, Examples, and Theorems by Stefan Gille

8 Complete fields and valuation rings

(1) is an invertible sheaf on X, which is generated by the global sections

Math 249B. Nilpotence of connected solvable groups

D-MATH Algebraic Geometry FS 2018 Prof. Emmanuel Kowalski. Solutions Sheet 1. Classical Varieties

FILTERED RINGS AND MODULES. GRADINGS AND COMPLETIONS.

14. Rational maps It is often the case that we are given a variety X and a morphism defined on an open subset U of X. As open sets in the Zariski

CHAPTER 7. Connectedness

CHAPTER 1. AFFINE ALGEBRAIC VARIETIES

Inseparable local uniformization. Conference on valuation theory, El Escorial

where m is the maximal ideal of O X,p. Note that m/m 2 is a vector space. Suppose that we are given a morphism

COMPLEX VARIETIES AND THE ANALYTIC TOPOLOGY

Extension theorems for homomorphisms

(1) A frac = b : a, b A, b 0. We can define addition and multiplication of fractions as we normally would. a b + c d

NONSINGULAR CURVES BRIAN OSSERMAN

12. Hilbert Polynomials and Bézout s Theorem

APPENDIX 3: AN OVERVIEW OF CHOW GROUPS

Math 203, Solution Set 4.

Chapter 8. P-adic numbers. 8.1 Absolute values

LECTURE 7: STABLE RATIONALITY AND DECOMPOSITION OF THE DIAGONAL

ALGEBRA EXERCISES, PhD EXAMINATION LEVEL

Math 249B. Geometric Bruhat decomposition

9. Integral Ring Extensions

Introduction to Arithmetic Geometry Fall 2013 Lecture #15 10/29/2013

On Maps Taking Lines to Plane Curves

3.1. Derivations. Let A be a commutative k-algebra. Let M be a left A-module. A derivation of A in M is a linear map D : A M such that

Definitions. Notations. Injective, Surjective and Bijective. Divides. Cartesian Product. Relations. Equivalence Relations

CHAPTER 0 PRELIMINARY MATERIAL. Paul Vojta. University of California, Berkeley. 18 February 1998

DIVISORS ON NONSINGULAR CURVES

(dim Z j dim Z j 1 ) 1 j i

MATH 8254 ALGEBRAIC GEOMETRY HOMEWORK 1

Math 120 HW 9 Solutions

9. Birational Maps and Blowing Up

Generalized Veronesean embeddings of projective spaces, Part II. The lax case.

Introduction to Algebraic Geometry. Jilong Tong

Subrings of Finite Commutative Rings

Isogeny invariance of the BSD conjecture

2. Intersection Multiplicities

ALGEBRAIC GEOMETRY COURSE NOTES, LECTURE 2: HILBERT S NULLSTELLENSATZ.

0 Sets and Induction. Sets

Projective Varieties. Chapter Projective Space and Algebraic Sets

Introduction to Arithmetic Geometry Fall 2013 Lecture #18 11/07/2013

arxiv: v2 [math.ag] 24 Jun 2015

Math 203A, Solution Set 6.

Honors Algebra 4, MATH 371 Winter 2010 Assignment 4 Due Wednesday, February 17 at 08:35

Thus, the integral closure A i of A in F i is a finitely generated (and torsion-free) A-module. It is not a priori clear if the A i s are locally

MATH 233B, FLATNESS AND SMOOTHNESS.

Math 248B. Applications of base change for coherent cohomology

IN POSITIVE CHARACTERISTICS: 3. Modular varieties with Hecke symmetries. 7. Foliation and a conjecture of Oort

Notas de Aula Grupos Profinitos. Martino Garonzi. Universidade de Brasília. Primeiro semestre 2018

Valuations and local uniformization

AN EXPOSITION OF THE RIEMANN ROCH THEOREM FOR CURVES

3. The Sheaf of Regular Functions

Math 594. Solutions 5

ne varieties (continued)

Algebraic Varieties. Notes by Mateusz Micha lek for the lecture on April 17, 2018, in the IMPRS Ringvorlesung Introduction to Nonlinear Algebra

Dedekind Domains. Mathematics 601

Resolution of Singularities in Algebraic Varieties

ON A THEOREM OF CAMPANA AND PĂUN

p-adic fields Chapter 7

Smooth morphisms. Peter Bruin 21 February 2007

LINEAR EQUATIONS WITH UNKNOWNS FROM A MULTIPLICATIVE GROUP IN A FUNCTION FIELD. To Professor Wolfgang Schmidt on his 75th birthday

Theorem 5.3. Let E/F, E = F (u), be a simple field extension. Then u is algebraic if and only if E/F is finite. In this case, [E : F ] = deg f u.

Introduction to Arithmetic Geometry Fall 2013 Lecture #17 11/05/2013

HYPERSURFACES IN PROJECTIVE SCHEMES AND A MOVING LEMMA

Clearly C B, for every C Q. Suppose that we may find v 1, v 2,..., v n

The Kronecker-Weber Theorem

Transcription:

HOMOMORPHISMS OF MULTIPLICATIVE GROUPS OF FIELDS PRESERVING ALGEBRAIC DEPENDENCE FEDOR BOGOMOLOV, MARAT ROVINSKY, AND YURI TSCHINKEL Abstract. We study homomorphisms of multiplicative groups of fields preserving algebraic dependence and show that such homomorphisms give rise to valuations. Introduction In this paper we formulate and prove a version of the Grothendieck section conjecture. For function fields of algebraic varieties over algebraically closed ground fields, this conjecture states, roughly, that the existence of group-theoretic sections of homomorphisms of their absolute Galois groups implies existence of geometric sections of morphisms of models of these fields. In detail, let k be an algebraically closed field, X an irreducible algebraic variety over k, and K = k(x) its function field. Let G K be the absolute Galois group of K. Fix a prime l not equal to the characteristic of k and let G K be the maximal pro-l-quotient of G K, the Galois group of the maximal l-extension of K. Write G a K = G K /[G K, G K ] and G c K := G K /[G K, [G K, G K ]], for the abelianization and its canonical central extension: (1) 1 Z K G c K π a G a K 1. Let Σ K = Σ(GK c ) be the set of topologically noncyclic subgroups σ GK a whose preimages π 1 a (σ) GK c are abelian. It is known that function fields K = k(x) of transcendence degree 2 over k = F p are determined, modulo purely inseparable extensions, by the pair (GK a, Σ K) [5], [7], and [14]. This raises the question of functoriality, i.e., the reconstruction of rational morphisms between algebraic varieties from continuous homomorphisms of absolute Galois groups of their function fields. This general fundamental question was proposed by Grothendieck and lies at the core of the Anabelian Geometry Program. Key words and phrases. Field theory, valuations. 1

2 FEDOR BOGOMOLOV, MARAT ROVINSKY, AND YURI TSCHINKEL The main open problem in this program relates to a Galois-theoretic criterium for the existence of rational sections of fibrations. Let π : X Y, be a fibration of integral algebraic varieties over k with geometrically irreducible generic fiber of dimension at least 1 over a base Y of dimension 2. This defines a field embedding π : k(y ) k(x), with the image of L := k(y ) algebraically closed in K := k(x). Dually, we have a surjective homomorphism of absolute Galois groups (a restriction map) G K G L, as well as the induced homomorphisms G c K G c L, G a K G a L. A minimalistic version of Grothendieck s Section conjecture, over algebraically closed k, would be: Conjecture 1. Assume that π a : G a K Ga L (2) ξ a : G a L G a K such that (3) ξ a (Σ L ) Σ K. admits a section Then there exist a finite purely inseparable extension ι : L L = k(y ) and a rational map ξ : Y X, such that ξ π (L) = ι (L) L. Thus ξ(y ) is a section over Y, modulo purely inseparable extensions. Conjecture 1 is closely related to questions considered in this note. Recall that, by Kummer theory, G a K = Hom(K, Z l (1)), and that (2) induces the dual homomorphism of pro-l-completions of the multiplicative groups ˆψ : ˆK ˆL. Then (3) says that ˆψ respects the skew-symmetric pairings on ˆK and ˆL, with values in the second Galois cohomology group of the

HOMOMORPHISMS AND ALGEBRAIC DEPENDENCE 3 corresponding field (with l-torsion coefficients). The groups ˆK and ˆL contain K /k and L /k, respectively. If the restriction ψ of ˆψ to K /k satisfies ψ : K /k L /k ˆL, then ψ respects algebraic dependence, mapping algebraically dependent elements in K to algebraically dependent elements of L (modulo k ). For function fields this is equivalent to (3) (see, e.g., [6, Section 5]). This relates the minimalistic version of the Section conjecture for rational maps to our main result, which we now explain. From now on, let K be an arbitrary field over k. Let ν be a nonarchimedean valuation of K, i.e., a homomorphism ν : K Γ ν onto a totally ordered group such that the induced map ν : K Γ ν { }, ν(0) =, satisfies a nonarchimedean triangle inequality. Let m K,ν o K,ν, K ν := o K,ν /m K,ν, o K,ν, be the maximal ideal, valuation ring, residue field, and units with respect to ν, respectively. If K k is a (transcendental) field extension and ν a valuation of K, then its restriction to k is also a valuation; and we have o K,ν k = o k,ν, o K,ν /o k,ν K /k, and a natural surjection We consider extensions of fields o K,ν /o k,ν K ν /k ν. k k k a K, where k is the prime subfield of K, i.e., k = F p or Q, and k a K the algebraic closure of k in K, i.e., the set of all algebraic elements over k contained in K. Assume that x 1, x 2 K /k satisfy (4) tr deg k( k(x 1, x 2 )) 1, for their lifts x 1, x 2 K ; and this does not depend on the choice of lifts. We write x 1 k x 2 and say that x 1 and x 2 are contained in the same one-dimensional field; clearly 1 k x, for all x K /k. From now on, we use the same notation for an element x K and its image in K /k. Let l l l a L

4 FEDOR BOGOMOLOV, MARAT ROVINSKY, AND YURI TSCHINKEL be field extensions, where l is the prime subfield of L, l a the algebraic closure of l in L, and let ψ : K /k L / l be a homomorphism of multiplicative groups. We say that ψ preserves algebraic dependence with respect to k, l if x 1 k x 2 ψ(x 1 ) l ψ(x 2 ), x 1, x 2 K /k. Theorem 2. Let k k K and l l L be field extensions as above. Assume that l = l a and that there exists a homomorphism (5) ψ : K /k L / l, such that ψ preserves algebraic dependence with respect to k and l; there exist y 1, y 2 ψ(k /k ), such that y 1 l y 2 ; ψ satisfies Assumption (AD) of Section 7. Then either (P) there exists a field F K such that ψ factors through K /k K /F, (V) there exists a nontrivial valuation ν on K such that the restriction of ψ to o K,ν /o k,ν K /k is trivial on (1 + m ν ) /o k,ν and it factors through the reduction map o K,ν /o k,ν K ν /k ν L / l, (VP) there exist a nontrivial valuation ν on K and a field F ν K ν such that the restriction of ψ to o K,ν /o k,ν factors through o K,ν /o k,ν K ν /F ν L / l. Note that we do not assume that k is algebraically closed. In the geometric setting treated in [8], when K = k(x) is a function field of an algebraic variety X over k = F p, case (P) corresponds to projections, the center of the valuation ν arising in case (V) is, birationally, the image of the section, and the above theorem can be viewed as a rational version of the minimalistic section conjecture (case (VP) corresponds to valuations composed with projections).

HOMOMORPHISMS AND ALGEBRAIC DEPENDENCE 5 To see this connection in more detail, recall that K admits a natural homomorphism with dense image to ˆK, the dual to GK a. The usual form of the section conjecture, as in Conjecture 1, is equivalent to the statement about homomorphisms ψ : ˆK ˆL such that pairs ˆf, ĝ with ( ˆf, ĝ) = 0 map to a pairs with (ψ( ˆf), ψ(ĝ)) = 0. Note that the image of K ˆK plays the role of a Z-sublattice, in the geometric case of function fields of algebraic varieties over algebraically closed fields. The corresponding statement of the theorem in the case of a function field is indeed a rational version of the section conjecture; we expect that it can be deduced from our version, using the fact that the natural sublattice K ˆK is necessarily mapped into L ˆL, modulo multiplication by a constant a Z l (as in [5]). Note that the abelian-by-central version of section conjecture does not hold for big fields. Indeed, as it was pointed in [3], the maximal extensions coprime to l of function fields have isomorphic Galois group, depending only on the algebraically closed ground field and the dimension. However, our theorem says that we can still obtain valuations from the multiplicative group homomorphisms respecting algebraic dependence. Here we extend the argument in [8] from function fields to arbitrary fields, under the additional technical assumption (AD) on ψ, which holds for K of positive characteristic. Although we believe that the main theorem holds in full generality, i.e., without the (AD) assumption, we were forced to add it, due to purely technical difficulties in our treatment of valuations of K which extend p-valuations of Q. To achieve clarity of the presentation, we decided to remove such valuations from present considerations. Related results on connections between Galois groups, valuations, and projective geometry can be found in [1], [9], [10], [11], [13]. The idea of the proof is to reduce the problem to a question in plane projective geometry over the prime subfield k, as in [4] and [5]. We view P(K) := K /k as a projective space over k. To establish Theorem 2, it suffices to show the existence of a subgroup U K /k such that: Condition 3. For every projective line l P(K), U l is either (1) the line l, (2) a point q l, (3) the affine line l \ q, or (4) if k = Q, a set projectively equivalent to Z (p) A 1 (Q) P 1 (Q),

6 FEDOR BOGOMOLOV, MARAT ROVINSKY, AND YURI TSCHINKEL the set of rational numbers with the denominator coprime to p. Indeed, such a subgroup is necessarily either F /k for some subfield F K, or o K,ν, for some valuation ν (see Section 7). By construction, the homomorphism ψ will satisfy the cases (P) or (V) in Theorem 2, respectively. To find such U, we use the results of [12] and [2]. First we deduce that the restriction of ψ to every plane P 2 P(K) is either an embedding or is induced by a natural construction from some nonarchimedean valuation (see Section 5). We distinguish two cases: there exists a line l P(K) such that the restriction of ψ to l is injective, no such lines exist. In the first case, property (4) of Condition 3 does not occur, and the proof works uniformly for k = F p or Q. In this case, we construct U by first taking the union u of all lines l(1, x) on which ψ is injective and then putting U := u u. We should that U is closed under multiplication and its intersections with projective subspaces Π P k (K) define a flag structure on Π, and thus a valuation on K. In the second case, the proofs are slightly different, leading to a case-by-case analysis in Section 5. Acknowledgments. We are very grateful to the referees for their careful reading of the first version of the paper, which helped us to improve the exposition. This work was partially supported by Laboratory of Mirror Symmetry NRU HSE, RF grant 14.641.31.0001. The first and second authors were funded by the Russian Academic Excellence Project 5-100. The first author was also supported by a Simons Fellowship and by EPSRC program grant EP/M024830. The third author was partially supported by NSF grant 1601912. 1. Projective geometry Let P be a projective space over a field k, i.e., the projectivization of a vector space over k. Let Π(q 0,..., q n ) P the projective envelope of points q 0,..., q n P. Working with lines and planes, we write l = l(q 0, q 1 ), resp. Π = Π(q 0, q 1, a 2 ), for a projective line through q 0, q 1, or a plane through q 0, q 1, q 2. Let ν a nonarchimedean valuation of k, o = o ν the corresponding valuation ring, and k ν the residue field. Fixing a lattice Λ o n+1 k n+1,

HOMOMORPHISMS AND ALGEBRAIC DEPENDENCE 7 we obtain a natural surjection (6) ρ = ρ Λ : P n (k) P n (k ν ). A 3-coloring of P 2 (k) is a surjection (7) c : P 2 (k) {,, }, onto a set of 3 elements, such that every l P 2 (k) is colored in exactly two colors, i.e., c(l) consists of two elements. A 3-coloring is called trivial of type I: if there exists a line l P 2 such that c is constant on P 2 \ l, II: if there exists a point q P 2 (k) such that for every l P 2 containing q, c is constant on l \ q. It was discovered in the early 1980s that such colorings are related to valuations, see, e.g., [12]. The same structure resurfaced in the study of the commuting elements of Galois groups of function fields in [4], exhibiting unexpected projective structures within GK a. This was a crucial step in the recognition of inertia and decomposition subgroups in GK a. Precisely, we have (see [12, Theorem 2] and [4]): Proposition 4. Assume that P 2 (k) carries a 3-coloring. Then there exists a nonarchimedean valuation ν such that the coloring c in (7) is induced from a trivial covering for some ρ as in (6). c ν : P 2 (k ν ) {,, }, 2. Flag maps We will consider maps (respectively, homomorphisms) f : P A from projective spaces over k to a set (respectively, an abelian group). The map f is called a flag map if its restriction f Π to every finite dimensional projective subspace Π P is a flag map. For k = F p and f : P n (F p ) A, this means that there exists a flag of projective subspaces (8) P n P n 1... P 1 P 0 = q such that f is constant on P i (F p ) \ P i 1 (F p ), for all i = 1,..., n. For k = Q and f : P n (Q) A,

8 FEDOR BOGOMOLOV, MARAT ROVINSKY, AND YURI TSCHINKEL this means that either there is a flag as in (8) so that f is constant on P i (Q) \ P i 1 (Q), for all i = 1,..., n, or there exist a prime p, a surjection as in (6), and a flag map such that ρ = ρ Λ : P n (Q) P n (F p ) f : P n (F p ) A, f = f ρ. Proposition 5. [4, Theorem 6.3.4] Let f : P(K) = K /k A be a group homomorphism which is also a flag map. Then there exist a valuation ν of K and a homomorphism r : Γ ν A such that f factors through (9) K /k ν Γ ν r A. A map f on P n (k) that is a flag map on every hyperplane defines an ˆf : ˆP n A λ f gen (λ) on the dual space, by assigning to a projective hyperplane the generic value of f on this hyperplane, i.e., the constant value on the complement to a codimension one subspace of that hyperplane. Every map f : P 2 (F 2 ) {0, 1} has the property that its restriction f l to any line l P 2 (F 2 ) is a flag map, but not all such f are flag maps. The following theorem, generalizing results in [4, Section 2], shows that this is the only exception. Theorem 6. Let f : P n (k) A be a map such that f l is a flag map, for every l P n (k). Then f is a flag map, unless k = F 2 and P n contains a plane Π = P 2 such that f is not a flag map on Π. Proof. We proceed by induction, assuming that f is a flag map on every P n 1 P n, n 2. We separate the cases: ˆf is constant, ˆf takes at least two values.

HOMOMORPHISMS AND ALGEBRAIC DEPENDENCE 9 In the first case, let a be the generic value of f on hyperplanes and S P n be such that f(x) a, for x S. The projective span P(S) of S is a proper subspace of P n, of codimension at least 2. Indeed, consider a subset of distinct elements x i S, i = 1,..., n; it contains an element with generic value of f on the span P(x 1,..., x n ). Hence, by assumption, dim P(x 1,..., x n ) n 2. Thus the dimension of P(S) is also bounded by (n 2). In particular, f(x) = a, for all x P n \ P(S). By induction, f is flag on P(S), thus also on P n. In the second case, let a λ := f gen (λ) be the generic value of f on Π λ P n, λ ˆP n. We have two possibilities: (1) There is a λ 1, with f gen (λ 1 ) = a 1, so that for q P n \ Π λ1, one has f(q) a 1. (2) There are λ 1 λ 2, with different generic values a 1, a 2, such that there are points q i P n \ Π λi with f(q i ) = a i. In Case (1), f is constant outside of Π λ1 and hence a flag map, by induction. Indeed, let P r Π λ1 be a projective subspace with f = a 1 on Π λ1 \ P r. On any line l(x 1, x 2 ) P n, with x 1, x 2 / Π λ1, intersecting Π λ1 \ P r at some point z 1, f is constant on l(x 1, x 2 ) \ z 1. Hence f(x 1 ) = f(x 2 ) in this case. It remains to show that for x i P n \ Π λ1 with z 1 P r we also have f(x 1 ) = f(x 2 ). Consider the Π(x, z 1, q), for any q Π λ1 \ P r i. If z 2 l(z 1, q), z 2 q, z 1, then l(x, z 1 ) and l(x 1, z 2 ) intersect at some point w P n \ Π λ1 and hence f(x 1 ) = f(w) = f(x 2 ), which concludes the proof. Case (2) does not occur unless n = 2. Indeed, let Π x1,q 1 be a hyperplane containing q 1 and intersecting Π λ1 \ P r nontrivially. The latter contains an affine plane A n 2 x in the intersection P1,x n 2 = P n 1 x,q 1 P1 n 1 and A n 2 x spans P n 2 1,x. Thus Px n 1 a 1 is the generic value on P n 1 is spanned by q 1 and A n 2 x and hence x,q 1. These hyperplanes are parametrized n 1 by ˆP 1 ˆP n minus a subset of hyperplanes which do not intersect Π λ1 \ P r. This set is empty if r < n 2 and consists of one point p 1 if r = n 2. Applying the same argument to a 2 we obtain a different ˆP 2 ˆP n n 1 n 1 n 1. The hyperplanes ˆP 1, ˆP 1 intersect at ˆP n 2 and we obtain a contradiction if n 2 > 0, since ˆP n 2 contains at least 3-points. Thus we obtain hyperplanes in P n with two generic values, contradicting the inductive assumption. We have reduced to n = 2, with the additional assumption that f is nonconstant on any line. Lemma 7. Case (2) does not occur for n = 2 unless k = F 2. Proof. Consider Π λ1 and its subset of generic points, which contains A 1 λ 1. Any line from q 1 to a point in A 1 λ 1 has generic value a 1. Let E 1

10 FEDOR BOGOMOLOV, MARAT ROVINSKY, AND YURI TSCHINKEL be the union of all such lines and E(a 1 ) E 1 the subset of points with value a 1. We define E 2 and E 2 (a 2 ) in a similar fashion. If k = F q then each set E 1 (a 1 ) and E 2 (a 2 ) contains at least q(q 1) + 1 points. The total number of points in P 2 (F q ) is (q 2 + q + 1), hence 2(q(q 1) + 1) (q 2 + q + 1) which implies q 2. It completes the proof for finite fields k. For infinite k, let l 1, l 2 be lines through q 2 E 2 which do not pass through q 1 and which intersect A 1 λ 1. These lines intersect all lines through q 1 E 1. Note that all those intersection points in E 1 \ Π λ1 are different for l 1 l 2. The generic value on l 1, l 2 is a 2 and hence at all but at most 4 lines l in E 1 through q 1 we have at least 2 points l l i with value a 2 which contradicts the fact that generic value on l is a 1. This concludes the proof of Theorem 6. Corollary 8. (1) Theorem 6 holds also P n (F 2 ) if f takes at least (n + 1) distinct values. (2) There is only one two-valued map on P 2 (F 2 ), modulo projective transformations, which is not a flag map. Proof. The first statement follows by induction on dimension, the case of P 2 (F 2 ) clear by Lemma 7. The second statement follows by direct verification. Lemma 9. Let f : P(K) = K /k A be a homomorphism whose restriction to every line is a flag map and such that there exists a plane Π = Π(1, x, y), with f(x), f(y) 1, and f Π not a flag map. Then f(x) = f(y) and f(x) 2 = 1. In particular, if f is not a flag map, then k = F 2 and f 2 is a flag map. Proof. Let Π := Π(1, x, y) P(K) such that f Π is not a flag map. Changing x, y by projective transformations and division by an element we can assume that f takes two values 1, a on Π, with f(1) = f(x + 1) = f(y + 1) = f(x + y) = 1 and f(x) = f(y) = f(x + y + 1) = a. On l(xy, x + y + 1), we have f(xy) = a 2, f(xy + x + y + 1) = 1, f(x + y + 1) = a

HOMOMORPHISMS AND ALGEBRAIC DEPENDENCE 11 and hence three values. Since a 2 a, by assumption, f is not flag on l(xy, x + y + 1). Lemma 10. Assume that the two-torsion part A[2] A is nontrivial. Consider the composition f 2 : P(K) f A r 2 A/A[2], with r 2 the projection. Then f 2 is a flag map on every plane Π P(K). Proof. If f is a flag map on Π(1, x, y), then f 2 is also flag. If f is not a flag map, then we apply Lemma 9: f takes only two values, 0 or 1, and f(x) 2 = 1, thus f(x) = 1. In particular, f 2 1 on Π(1, x, y) and hence is a flag map. Thus f 2 is a flag map on every plane, and hence a flag map. To summarize, if A A[2] then f determines a valuation ν. If A = A[2], then f is trivial on some subfield K K such that K K is a purely inseparable extension of exponent 2. 3. Z (p) -lattices Let p be a prime number and Z (p) Q the set of rational numbers with denominator coprime to p. A Z (p) -lattice, or simply, a lattice B Q n+1 is a Z (p) -submodule such that B Z(p) Q = Q n+1. Given a lattice B Q n+1 and an element x Q n+1 \ 0, there exists an element x B B \ pb such that x and x B define the same point in P n (Q), this element is unique in B \ pb, modulo scalar multiplication by Z (p). Lattices B, B Q n+1 are called equivalent if B = a B, for some a Q. In this section, we consider the maps f : (Q n+1 \ 0) A, which are invariant under scalar multiplication by Q ; and we use the same notation for the induced map f : P n (Q) A. We say that f is induced from P n (Z/p) via a lattice B if there exists a map f : P n (Z/p) A, such that where f(x) = ( f ρ B )(x B ), for all x P n (Q), ρ B : (B \ pb) (B/pB) \ 0 P n (Z/p).

12 FEDOR BOGOMOLOV, MARAT ROVINSKY, AND YURI TSCHINKEL This is well-defined since ρ B is invariant under Z (p). Any such lattice will be called f-compatible, or simply compatible. If f is induced from P n (Z/p) via a lattice B, then it is also induced via any equivalent lattice. Any sublattice of Q n is compatible with a constant function. However, if f takes at least two values then the set of f-compatible lattices is much smaller. Note that equivalence classes lattices in Q 2 are naturally parametrized by a p-tree (a tree where each vertex has (p + 1) outgoing edges). Lemma 11. If f is a Q -homogeneous function on Q 2 \ 0 and f(x) f(y) then the set of f-compatible lattices consists of Z (p) -lattices generated by p mx x, p my y, with m x, m y Z. Proof. Let B be f-compatible and consider the projection ρ B. Then, for x Q 2 \0 there is a unique m x Z such that p mx x is a generator of B, i.e., p mx x B \ pb. We have ρ B (p mx x) = x P 1 (B/pB). Consider y Q 2 \ 0 and p my y B \ pb with ρ B (p my y) = ȳ P 1 (B/pB). Then ρ B (p my y) ρ B (p mx x) since f is induced from f on P 1 (B/pB) and f(x) f(y). This implies that p mx x and p my y generate B. In the discussion below, we use projective and affine geometry. The following lemma connects these concepts. Lemma 12. Assume that f : P 1 (Q) A is induced from a nonconstant map f : P 1 (Z/p) A, via some lattice. (1) If f is a flag map, then there are exactly two equivalence classes of f-compatible lattices B 1, B 2 Q 2. (2) If f is not a flag map, then there is exactly one equivalence class of f-compatible lattices B Q 2. Proof. By assumption, f is induced via some ρ B. Fix generators x, y B such that f(y) f(x), in particular ρ B (x B ) ρ B (y B ) P 1 (Z/p). We have f(y + pb) = f(y) and f(x + pb) = f(x) f(y). Any lattice B Q 2 is equivalent to a lattice with x as a generator. Since B /Z (p) x Z (p), B is one of the following: B i := x, p i y, for some i Z. If f is induced from B i, for some i < 1, then f(x + p(p i y)) = f(x) f(y) and f(x + p(p i y)) = f(p i 1 x + y) = f(y), a contradiction. The same argument yields a contradiction when i > 1. Thus i = 1, 0, or 1. Analysis of values of f at other points of P 1 (Z/p) leads to further restrictions. We have the following cases:

HOMOMORPHISMS AND ALGEBRAIC DEPENDENCE 13 (1) f is constant on P 1 (Z/p) \ ρ B (y B ). (2) f is not constant on the complement to a point in P 1 (Z/p). In Case (1), f(x + y) = f(x), excluding i = 1. Then we have exactly two lattices B 0, B 1, such that f is induced from these (or equivalent) lattices. In Case (2), if f is induced from B 1 then f(κx + y) = f(pκx + p 1 y) = f(y), for any κ Z (p), and hence f is constant on P 1 (Z/p) \ ρ B (x B ), contradicting the second condition. A similar argument works for B 1. Thus there is only one compatible lattice B 0 = B, modulo equivalence. A similar analysis holds for f-compatible lattices in Q n, for arbitrary n. For x Q n \ 0 we let x be the ray consisting of its nonzero multiples. Then, for any sublattice B, the image of x B in P n 1 (Z/p) is well-defined. However, it may happen that for some B the corresponding images lie in a proper subspace of P n 1 (Z/p) while for another f-compatible lattice they span the whole P n 1 (Z/p). Lemma 13. Assume that f is induced from f via some lattice B and that the images x 1 B, x 2 B,..., x n B span P n 1 (Z/p). Then B is generated by p i 1 x 1,..., p in x n, for some i 1,..., i n. Lemma 14. Assume that f : P 2 (Q) A satisfies the following: (1) f takes three values; (2) f takes at most two values on every line l P 2 ; (3) on every P 1 (Q) P 2 (Q), f is induced from a flag map on P 1 (Z/p), via ρ B, for some lattice B Q 2. Then there are exactly three equivalence classes of lattices B i Q 3 such that f is induced from a flag map f : P 2 (Z/p) A, via ρ Bi, i = 1, 2, 3. Proof. It follows from Proposition 4, applied to k = Q (see also [12] or [4]). The first two conditions imply that there exists a lattice B Q 3 such that f is induced from some We conclude that f is a flag map, with 3 distinct values. Hence P 2 (Q) = S 1 S 2 S 3, with S 1 the preimage of an affine plane in P 2 (Z/p), S 2 an affine line, and S 3 a point in P 2 (Z/p), and f is constant on these sets. Thus, for any B Q 3 such that f is induced from P 2 (Z/p) via ρ B, the restriction of f to any (Q 2 \ 0) (Q 3 \ 0) is induced from a flag map on P 1 (Z/p). Hence f is also induced from a flag map, via ρ B. On the other hand, in coordinates x 1, x 2, x 3, we have S 1 = {x 1 0}, S 2 = {x 1 = 0, x 2 0}, S 3 = {x 1 = x 2 = 0, x 3 0},

14 FEDOR BOGOMOLOV, MARAT ROVINSKY, AND YURI TSCHINKEL and the only possible coordinates compatible with the structures on all P 1 (Q) are x 1, x 2 p, x 3 p, x 1, x 2, x 3 p, and x 1, x 2, x 3. Indeed, consider lattices B 1, B 2, B 3 as above and assume that there exists another sublattice B Q 3 such that f is induced via B. Choose Q 2 1 Q 3 so that f is not constant on Q 2 1. Then we can choose B i so that B i Q 2 1 = a B Q 2 1, a Q, since there are only two possible equivalence classes of compatible lattices in Q 2 1. Note that there are at least two such B i whose intersections with a complementary subspace Q 2 2, with f nontrivial on it, are different. Thus for at least one B i we can assume that B Q 2 j = B i Q 2 j, j = 1, 2 and B i Q 2 j generate B i over Z (p). Hence B is equal to B i, for one of the i = 1, 2, 3. Let 4. A result from field theory k k k a K be an extension of fields. We say that x 1, x 2 K /k are algebraically dependent with respect to k if they satisfy Equation (4) from the Introduction; in this case, we write x 1 k x 2, or simply x 1 x 2. We record the following obvious properties of this equivalence relation: (AI) If x 1 k x 2, x 1 /x 2 / k a /k, and x k x 1, then x 1 /x k x 2 /x. (AG) For all x 1 K /k \ k a /k, the set of x 2 such that the closure of k a (x 1 ) in K coincides with the closure of k a (x 2 ), together with ( k a /k ), forms a subgroup of K /k. Lemma 15. Let K k and L l be field extensions, ν a valuation of K, and (10) ψ : K /k L /l a homomorphism, such that its restriction to o K,ν /o k,ν (11) o K,ν /o k,ν K ν /k ν ψ ν L /l. factors as Assume that ψ ν preserves algebraic dependence with respect to k ν and l. Then ψ also preserves algebraic dependence with respect to k and l.

HOMOMORPHISMS AND ALGEBRAIC DEPENDENCE 15 Proof. Let k(x) K be a purely transcendental extension and E = k(x) K its algebraic closure in K. We claim that the restriction of ψ to E /k preserves algebraic dependence. There are two cases: Case 1. ν(k ) = ν(e ). Then E = o E,ν k. Since ψ ν preserves algebraic dependence with respect to k ν and l, the claim follows. Case 2. ν(k ) ν(e ). Then ν(e )/ν(k ) has Q-rank 1, i.e., for y, z E with nonzero ν(y), ν(z) ν(e )/ν(k ) there are nonzero n y, n z Z such that n y ν(y) = n z ν(z). Indeed, y, z define a finite algebraic extension k y,z (x) k(x), hence ν is nontrivial on k(x), and the group ν(k y,z (x) )/ν(k(x) ) is finite. Let g k(x) be such that the image of ν(g) in ν(e )/ν(k ) is infinite. Then for any n i=0 a ig i, with a i k, ν( n i=0 a i g i ) = min i (ν(a i g i )), since none of the monomials a i g i have the same value under ν. Thus, ν(k(g) ) = ν(k ) ν(g). The extensions k y,z (x) k(x) and k(x) k(g) are finite, thus ν(k y,z (x) )/(ν(k ) ν(g) ) is also finite, which implies the result for ν(e ). Since ψ(k ) = 1, ψ(k y,z (x) ) is the product of a finite group and Z. In particular, ψ(k y,z (x) ) consists of algebraically dependent elements. Since E is a union of subfields k y,z (x), the same holds for E. Thus ψ(e /k ) coincides with the image of ν(e )/ν(k ). Since all elements in ν(e )/ν(k ) have the same powers, i.e., the Q-rank of this group is at most one, we see that the lifts of elements in ψ(e ) to L are algebraically dependent over l.

16 FEDOR BOGOMOLOV, MARAT ROVINSKY, AND YURI TSCHINKEL 5. Restriction to planes Here we study the restrictions of homomorphisms ψ : P(K) = K /k A := L / l, satisfying the assumptions of Theorem 2, to projective planes Π P(K). Proposition 16. Let Π := Π(1, x, y) P(K) be a projective plane such that ψ(x) ψ(y). Then one of the following holds: (a) ψ Π is injective. (b) There exists a line l Π such that ψ Π is constant on Π \ l. (c) There exists a point q Π such that ψ Π is constant on l \ q, for every l Π passing through q. (d) k = Q, ψ Π is induced from ψ Π : P 2 (Z/p) A, via a lattice B Q 3, and ψ Π is of type (a), (b), or (c). Proof. Assume that ψ Π is not injective: there are distinct x 1, x 2 Π, with ψ(x 1 ) = ψ(x 2 ) 1. Consider Π 1 := x 1 1 Π = Π(1, 1/x 1, y/x 1 ), since ψ(y) ψ(1/x 1 ), Π 1 satisfies the conditions of the theorem; if it holds for Π 1, then it holds for the initial Π. Thus we may assume that (12) S 1 := {x Π ψ(x ) = 1} contains at least two elements. Consider the map ψ : P(K) A, with values in dependency classes: ψ (x ) = 1 if ψ(x ) = 1, ψ (x ) = ψ (x ) iff ψ(x ), ψ(x ) 1 and ψ(x ) ψ(x ). We record the properties of ψ : (TI) For every l Π with l S 1 =, we have {ψ (x ) x l} = {ψ (x ) x Π \ S 1 }, in particular, ψ(l) has algebraically independent elements. (TC) For every l Π with l S 1, ψ is constant on l \ (l S 1 ). Property (AI) from Section 4 relates ψ and ψ. Lemma 17. If l S 1 = and x, x l are such that ψ(x ) ψ(x ), then ψ(x ) = ψ(x ).

HOMOMORPHISMS AND ALGEBRAIC DEPENDENCE 17 Proof. There is a z l with ψ(z) ψ(x ), ψ(x ). Since z 1 l S 1, all values of ψ on l(x /z, x /z) \ 1 are algebraically dependent (here we use that l a = l). By (AI), if ψ(x ) ψ(x ), then ψ(x )/ψ(z) ψ(x )/ψ(z), a contradiction. Lemma 18. Let l, l Π be disjoint from S 1, put z := l l, and assume that there exist x l and x l such that ψ(x) ψ(x ), ψ(x) ψ(x ), and ψ(x), ψ(x ) ψ(z). Let y l and y l be such that ψ(y) ψ(y ). Then either ψ(y) ψ(y ), or ψ(y) = ψ(y ) = ψ(z). Proof. Assume that ψ(y) = ψ(y ) ψ(z), then ψ(x) ψ(z) ψ(y) ψ(z) and ψ(x ) ψ(z) ψ(y ) ψ(z), by (TC), by the same argument as in Lemma 17. Finally, by (AI), is not possible. ψ(x) ψ(z) ψ(x ) ψ(z) Let {T j } j J be the set of intersections of algebraic dependency classes in P(K) with Π. Split J = J 2 J 3 and consider the decomposition (13) Π = S 1 S 2 S 3, with S 1 = T 1, S 2 = j J2 T j, S 3 = j J3 T j, (here S 1 is the same set as in (12)). For any such decomposition, the induced map Ψ = Ψ Π : Π {1, 2, 3} factors through ψ and satisfies the conditions of Proposition 4. Thus Ψ is induced from a trivial coloring, with S 1 not depending on the decomposition. Since there exist lines disjoint from S 1, and S 1 contains at least two points, it follows that either (B) S 1 = Π \ l, for some l Π, and we are in Case (b), or (C) S 1 = i I (l i \ q), for some q Π and l i through q, and we are in Case (c), as is proved below in Lemma 19, or (D) k = Q, and Ψ is induced from a trivial coloring on P 2 (Z/p). Note that in Case (B), ψ 1 on the affine plane Π \ l.

18 FEDOR BOGOMOLOV, MARAT ROVINSKY, AND YURI TSCHINKEL Lemma 19. In case (C), ψ is constant on an affine plane, or on l \ q, for all lines l passing through q. Proof. Consider x Π \ (S 1 q) and lines l containing x but not q. Then ψ ψ (x) on l\(l S 1 ), by (TI). Since S 1 is not an affine plane, there is an x Π \ (S 1 l(x, q)). We have ψ (x) = ψ (x ). The union of lines l Π, q / l, through x, x, is equal to Π\q. Thus ψ takes only three values {1, ψ(x), ψ(q)} and is constant on Π \ (S 1 q). Lemma 17, applied to l through q, implies that ψ is constant on l \ q. We are left with Case (D), when Ψ is induced via some from a trivial coloring ρ : Π = P 2 (Q) P 2 (Z/p) c : P 2 (Z/p) {1, 2, 3}, in the sense of Proposition 4. Put S i = c 1 (i), i = 1, 2, 3. Note that S 1 is a finite union of subsets Z (p) +Z (p) and does not contain a complete line l. Consider shifts Π z := z 1 Π, for z Π. The shift from Π to Π z changes algebraically dependent subsets. Note that Π z contains 1 and S 1 contains Z (p) + Z (p), by assumption of case (D), thus on lines l through 1 all elements in l \ 1 are algebraically dependent. If there are at least two elements x/z, y/z with ψ(x/z) ψ(y/z) then we have a splitting into S 1, S 2, S 3, and since S 1 contains Z (p) + Z (p) we can proceed by induction. Lemma 20. For every z Π, the restriction of ψ to Π z is induced from P 2 (Z/p). Proof. We subdivide (D) into subcases: (D1) For every z and every splitting Π z = S 1,z S 2,z S 3,z, where S 2,z, S 3,z are unions of algebraic dependency classes, the set S 1,z P 2 (Z/p) is either a point, an affine line, or an affine plane. (D2) Otherwise: for some Π z this is not the case. First we treat (D1). Fix Π z and a decomposition Π z = S 1,z S 2,z S 3,z ; we have Ψ : P 2 (Z/p) {1, 2, 3}, and P 2 (Z/p) = 3 i=1 S i,z, S i,z = ρ 1 ( S i,z ) Π. By assumption (D1), we have 3 cases.

HOMOMORPHISMS AND ALGEBRAIC DEPENDENCE 19 S 1,z = q, for some q P 2 (Z/p). For x Π z \S 1,z and l = l(q, x), with ρ(q) = q, ψ is constant on l \ (l S 1,z ), by (TC). Apply this to all l(q, x 1 ), where x 1 runs over S 1,z, to conclude that ψ is constant on preimages of affine lines ( l \ q), with q l, hence is induced from P 2 (Z/p). S 1,z = l \ q, for some l P 2 (Z/p) and q l. Then S 1,z, S 2,z and S 3,z form a flag on P 2 (Z/p): all points projecting to P 2 (Z/p) \ l belong to the same algebraic dependency class because each pair of such points can be connected by a pair of lines which intersect S 1,z. Lemma 14 reduces the proof to the previous case, after changing to a different ψ -compatible lattice. S 1,z = P 2 (Z/p) \ l, for some line l P 2 (Z/p). This reduces to the case S 1,z = q. We pass to (D2) and fix a plane Π z, with a splitting Π z = S 1,z S 2,z S 3,z, violating (D1). Note that Π z contains points 1, 1/z, x/z with ψ(1/z) ψ(x/z). We also know that the subset S 1 /z Π z, with S 1 defined in (13), is a finite union of subsets projectively equivalent to Z (p) + Z (p) P 2 (Q). By Proposition 4, we have an induction of Ψ z from the trivial coloring on P 2 (Q ν ), for some valuation ν on Q, and we obtain that ν is nontrivial on Q, i.e., corresponds to some prime number. Since S 1 /z is contained in either S 2,z or S 3,z, the corresponding prime equals p. Thus Ψ z is induced from a trivial 3-coloring of P 2 (Z/p). Now, we may assume that S 1,z is a union of more than one subset of type l i \ q P 2 (Z/p) (otherwise, we are in Case (D1)). Note that one of the subsets S i,z, i = 2, 3 is q and the complement of all such is S 3,z. Then there exist a point q P 2 (Z/p) and a set { l i } i I of at least two lines passing through q such that S 1,z = i I ( l i \ q), since we are in the case (D2), by assumption. Moreover, we may assume that S 2,z = q, then S 3,z has the same structure as S 1,z, i.e., a union of affine lines containing q in their closure. We claim that ψ is constant on S 3,z : consider q 3, q 3 S 3,z not lying on a line through q. Let q 3, q 3 be any points projecting to q 3, q 3. Since l( q3, q 3) S 1, the line l(q 3, q 3) intersects S 1, thus ψ(q 3 ) ψ(q 3). By assumption on S 3,z, any two points in S 3,z can be connected by a chain of such lines. Note that ψ is constant on S 2,z : consider q 1, q 2 with ρ(q 1 ) = ρ(q 2 ) = q S 2,z. Then ψ(q 1 ) = ψ(q 2 ). Indeed, consider l 5 = l(q 1, x 1 ) and l 6 = l(q 2, x 2 ), where ρ(x i ) = x i S 1, x 1 x 2. Hence q 3 := l 5 l 6 projects to q. Thus

20 FEDOR BOGOMOLOV, MARAT ROVINSKY, AND YURI TSCHINKEL ψ (q 1 ) = ψ (q 3 ) = ψ (q 2 ). Thus ψ is constant on S 2,z, hence ψ is induced from P 2 (Z/p). Now we use ψ to prove the induction from P 2 (Z/p) result for ψ itself. The difference between S 1 and S 2, S 3 is that ψ is already constant on S 1 but not necessarily on S 2, S 3. We treat the cases: (1) S 1 = q, S 2 = l \ q, q l, S 3 = P 2 (Z/p) \ l; (2) S 1 = q, S 2 = m i=1 l i \ q, m 2, q l i, S 3 = P 2 (Z/p) \ l i ; (3) S 1 = m i=1 l i \ q, m 2, S 2 = q, S 3 = P 2 (Z/p) \ l i. Lemma 21. The map ψ Π is induced from ψ Π : P 2 (Z/p) A which is of the type (a), (b), or (c). Proof. By Lemma 20, we have the following possibilities: (1) ψ is induced from a flag map on P 2 (Z/p) and we can assume that S 1 = q, by Lemma 14; (2) ψ is induced from a map on P 2 (Z/p) which is constant on affine lines l i \ q, with q l, and S 1 = q; (3) ψ is induced from a map on P 2 (Z/p) which is constant on affine lines l i \ q, with q l, and S 1 contains l i \ q, i = 1, 2. Case (1): We may assume that S 3 = P 2 (Z/p) \ l, for some l with q l, and l \ q = S 2. Let l be disjoint from S 1 and pick two points q, q l S 3. Since ψ (q) = ψ (q ) and l intersects S 2, ψ(q) = ψ(q ), by Lemma 17. Since any two points in S 3 can be connected by a chain of lines disjoint from S 1, ψ is constant on S 3. It is also constant on ρ 1 ( q 2 ), for q 2 S 2. Indeed, if q 2, q 2 are distinct points projecting to q 2 and l, l lines containing q 2, resp. q 2, avoiding S 1 and projecting to distinct lines in P 2 (Z/p), then q 2 := l l also projects to q 2. Thus ψ(q 2 ) = ψ(q 2) = ψ(q 2). Case (2): S1 = q. If ψ is induced from a noninjective ψ : P 2 (Z/p) A, ψ is constant on the preimage of every affine line l \ q, by the same analysis over a finite field. If there exist y 1, y 2, projecting to the same points x l \ q, with ψ(y 1 ) ψ(y 2 ), let z 1, z 2 be such that ψ (z 1 ) = ψ (z 2 ) but ψ (z i ) ψ (y i ). Consider z := l(y 1, z 1 ) l(y 2, z 2 ), so that ρ(z) = x. Then ψ(y 1 ) = ψ(z) = ψ(y 2 ), by Lemma 18. Since all points over x are connected by a chain of lines of such type, ψ is constant on ρ 1 ( x).

HOMOMORPHISMS AND ALGEBRAIC DEPENDENCE 21 Case (3): The argument of Case (1) shows that ψ is constant on the preimage of any affine line l \ q contained in S 3. Indeed, let z 1, z 2 S 3 be in the preimage of an affine line S 3 and consider l := l(z 1, z 2 ). It intersects S 2 and hence ψ(z 1 ) = ψ(z 2 ). Thus ψ is induced from P 2 (Z/p) \ q = S 1 S 3. Let q, q, projecting q. Consider lines l(q, z 1 ) and l(q, z 2 ) with z i S 3, which intersect in q, ρ(q ) = q. Then ψ(q) = ψ(q ) = ψ(q ), by Lemma 18. Since any pair of points over q can be connected by a chain of such lines, ψ is constant on ρ 1 ( q). This concludes the proof of Proposition 16. Remark 22. This Lemma is similar to [12] and [8, Lemma 13]. 6. Lines of injectivity In our analysis of the restriction ψ l of ψ : P(K) A = L / l to lines l = l(1, x) P(K), we distinguish the following possibilities: ψ l is not induced from a map ψ l : P 1 (Z/p) A and ψ l is: (I) injective (N) not injective and nonflag (F) a nonconstant flag map ψ l is induced from ψ l : P 1 (Z/p) A and ψ l is (Ī) injective ( N) not injective and nonflag ( F) a nonconstant flag map (C) ψ l is constant Definition 23. Let u P(K) be the union of all lines through 1, on which ψ is injective, and put U := {xy x, y u} P(K). Lemma 24. If ψ(u) contains at least two algebraically independent elements, then U is a group. Proof. Clearly, u and U contain 1 K /k. If x U then x 1 U, by the injectivity of ψ on l(1, x 1 ). Furthermore, (14) xy 1 u, for all x, y u such that ψ(x) ψ(y). Indeed, if ψ(x) ψ(y), then ψ is injective on Π(1, x, y), by Proposition 16, and in particular on l(x, y) = y l(1, xy 1 ); thus, xy 1 u. If ψ(x) ψ(y), but are not equal in A, take z u such that ψ(x) ψ(z). Then x/z, y/z u, as above. Since ψ(x/z) ψ(y/z), the same argument shows that (x/z)/(y/z) = xy 1 u, proving (14).

22 FEDOR BOGOMOLOV, MARAT ROVINSKY, AND YURI TSCHINKEL To show that U is multiplicatively closed, it suffices to check that for every x 1, x 2, x 3 u \ {1} there exist s 1, s 2 u with x 1 x 2 x 3 = s 1 s 2. Note that ψ(x i x j ) 1 for some 1 i < j 3. (Otherwise, and therefore, ψ(x 1 x 2 ) = ψ(x 1 x 3 ) = ψ(x 2 x 3 ) = 1, ψ(x 2 1) = ψ((x 1 x 2 )(x 1 x 3 )/(x 2 x 3 )) = 1, so ψ(x 1 ) = 1.) Then, by (14), x i x j u, so we can take s 1 := x i x j and s 2 := x t, where {i, j, t} = {1, 2, 3}. Definition 25. Let ū P(K) be the union of all lines l through 1, such that the restriction of ψ to l is induced via an injective map ψ l : P 1 (Z/p) A, and put Ū := {xy x, y ū} P(K). Lemma 26. If ψ(ū) contains at least two algebraically independent elements, then Ū is a group. Proof. The proof follows the same steps as the proof of Lemma 24. Lemma 27. Assume P(K) contains lines of type (I) and one of the types (15) (N), (Ī), ( N), or ( F). Then there exists a one-dimensional subfield E L such that for all lines l P(K) of type (I), (N), (Ī), ( N), or ( F) we have ψ(l) E / l. In particular, if ψ(u) contains algebraically independent elements, lines of type (N), (Ī), ( N), and ( F) do not exist. Proof. Let l = l(1, y) be a line of type (I). If there exists another line l(1, y ) of type (I) with ψ(y) ψ(y ), i.e., ψ(u) contains independent elements, then lines of the listed type cannot exist, indeed, if l(1, x) is of types listed in (15), we apply Proposition 16 to Π = Π(1, x, y). In Case (b), the exceptional line is l(1, y) and hence the restriction of ψ to any other line is either constant or of type (F), a contradiction. In Case (c), all lines are either of type (I) or (F), again a contradiction. Case (d) does not apply, since l(1, y) is not induced from a map P 2 (Z/p) A, a contradiction. If ψ(u) does not contain algebraically independent elements, but one of the lines l(1, x) in (15) is such that ψ(x) ψ(y), then we apply the same argument to Π(1, x, y) and obtain the same contradiction.

HOMOMORPHISMS AND ALGEBRAIC DEPENDENCE 23 Lemma 28. Assume ψ(u) contains algebraically independent elements. Consider l := l(1, y) u and assume that l U consists of at least two points 1, z. Then l U is either l or l \ q, for some point q l. Proof. Assume that ψ l is not constant, e.g., ψ(y) 1. By assumption, there is an x with l(1, x) u with ψ(x) ψ(y). We apply Proposition 16 to Π := Π(1, x, y). We are not in Case (c) of this lemma. If we are in Case (a), then ψ is constant on Π \ l(1, x), which implies that l is of type (F). If we are in Case (b), then the exceptional point q = y, and ψ is constant, on the complement to q, on every line through q, thus l is of type (F). Put z = t/t, with t, t u. If ψ(t) ψ(t ) then Equation (14) implies that z u, a contradiction. Thus ψ l is either constant or contains one point y / U. In Case (a), ψ is constant on Π \ l(1, x), thus identically 1 on the line l. In Case (b), ψ is injective on every line not containing the exceptional point q, in particular on l(1, t /t ), for all t, thus t /t u, thus t U. Taking t l \ q we obtain the claim. Now assume that ψ l is constant. We claim that l \ (l U) contains at most one point. Assume otherwise. Note that ψ is injective on every line l(u, t ) Π, with t Π(1, x, y) u, t 1, and any point u l U. Indeed, we can represent u = w/w, with w, w u and with ψ(w) = ψ(w ) ψ(t ). Then t w /w u and l(t w /w, 1) u. The converse is also true, and (Π \ l) u. Indeed, consider lines through u which are not equal to l; ψ is injective on such lines. Now consider two families of lines: those passing through w (except l), and those throgh w (again, except l). All such lines are of type (F), with generic value 1, since ψ does not take value 1 on Π \ l, by Lemma 27. Consider lines l(w, v) and l(w, v) from these families, with v (Π \ l). The generic ψ-value on these lines is the same and is equal to ψ(v). A line through u, which does not contain v cannot be of type (I), since it intersects lines l(w, v) and l(w, v) in distinct points, but taking the same value on these points, contradicting the established fact that such lines are of type (I). Lemma 29. Assume P(K) contains lines of type (Ī) and there exist lines of type (I), or (N), or ( N). Then ψ(ū) does not contain algebraically independent elements. Proof. Assume the contrary. Let l(1, x) be a line of type (I) or (N). Then there exists a y ū such that ψ(y) ψ(x). We apply Proposition 16 to Π = Π(1, x, y) and obtain a contradiction as in the proof of Lemma 27.

24 FEDOR BOGOMOLOV, MARAT ROVINSKY, AND YURI TSCHINKEL Let l(1, x) be of type ( N). We claim that Π does not contain lines of type (F). To exclude this possibility, let l = l(z, t) Π be such a line with a generic ψ-value equal to s A. Take points x 1, x 2 l(1, x) such that ψ(x 1 ) ψ(x 2 ). This is possible since ψ takes at least two values on l(1, x). Choose y 1, y 2 l(1, y) and ψ(y 1 ) ψ(y 2 ) and are both not equal to 1 A, and this is possible because ψ takes at least three values on l(1, y) which is of type (Ī). y 2 q y 1 1 x 1 x 2 Moreover, we can assume that the lines l ij := l(x i, y j ) do not pass through the distinguished point q l(z, t) (where ψ takes the nongeneric value). Thus l ij := l ij l(z, t) is a generic point of l(z, t), which differs from x 1, x 2, y 1, y 2. Then for both i = 1, 2. Hence ψ(x i ) s ψ(y 1) s ψ(y 2), s 1 ψ(x 1) ψ(x 2 ) ψ(y 1) ψ(y 2 ) 1. Therefore, ψ(x) ψ(y), a contradiction. Thus, for every l Π(1, x, y) the restriction ψ l is induced from a map ψ l : P 1 (Z/p) A. Now we apply Lemma 21. In Case (a) of that Lemma, the exceptional line is l(1, y) and of type (Ī) and hence the restriction of ψ to any other line is either constant or of type ( F), a contradiction the assumption that l(1, x) is of type ( N). Cases (b) and (c) are excluded: ψ is not induced from an injective map, nor a flag map on l(1, x). Lemma 30. Assume that the pair of lines (l(1, x), l(1, y)) is of one of the following types Then ψ(x) ψ(y). (N, N), (N, N), (N, F), ( N, N). Proof. Follows from the same arguments as in Lemma 29 and Lemma 27.

HOMOMORPHISMS AND ALGEBRAIC DEPENDENCE 25 Lemma 31. Assume that ψ(ū) contains algebraically independent elements. Consider l := l(1, z) ū, and assume that l Ū consists of at least two points 1, z. Then ψ(z ) = 1 and l Ū is either (1) l; (2) an affine line, with ψ not constant on l; (3) projectively equivalent to Z (p) P 1 (Q); (4) an affine line and ψ is constant on l. Proof. Assume that l / (C). Write z = x/x with x, x ū. If ψ(x) ψ(x ) then x/x ū, by Equation 14, thus l ū Ū, contradiction, so that ψ(z ) = 1. If ψ(x) = ψ(x ) 1, choose a point t ū be such that ψ(t) ψ(x ), and such that it is also algebraically independent from a nontrivial value on l. Then t/x ū and the restriction of ψ to (a shift of) l(t, x/x ) is of type (Ī). In particular, l(1, t), l(t, x/x ) are also of type (Ī), by the same argument as in the proof of Lemma 29. This lemma implies that l is of type (F) or ( F). l (F). In the notation of Proposition 16, ψ is of type b) on Π(1, t, z) and the restriction of ψ to every line in Π(1, t, z), not passing through a distinguished point q l, with ψ(q) 1, is of type (Ī), which implies that l \ q U, i.e., we are in Case (V), i.e., the valuation case. l ( F). In this case, Π(1, t, z) does not contain lines of type (F), because otherwise, by Proposition 16, l will also be of type (F). Hence ψ is induced from P 1 (Z/p) on any line in Π(1, t, z) and there are two independent values of ψ on Π(1, t, z) not equal to 1. Then ψ on Π(1, t, z) is induced from ψ : P 2 (Z/p) A, by Lemma 21. The map ψ is injective on l(1, t ) and l(t, z 1), where both t, z 1 are the images of t, z under the reduction map, and a flag map on l(1, z 1 ), where z 1 is the image of z in P 2 (Z/p). Thus ψ is induced from type b), and hence U l consists of y, with ψ(y) = 1, a set projectively equivalent to Z (p) P 1 (Q), and we are in Case (P), the projection case. Assume that l (C). Here the difficulty is that ψ(π(1, z, t)) does not contain algebraically independent elements and we cannot apply Lemma 21. Note that l(t, s), for s = r/r, r, r u, s l, are of type (Ī), by the argument above.

26 FEDOR BOGOMOLOV, MARAT ROVINSKY, AND YURI TSCHINKEL Then any line l(t, u ) Π(1, z, t), with ψ(t ) ψ(u ), is of type (Ī), since ψ takes at least three values on this line. Hence s := l(t, u ) l Ū. On the other hand, if s l is not in Ū, then there are at most two values on any line containing s, including ψ(s ) = 1. We split all points into subsets: (1) S T := {x ψ(x) 1}; (2) S 1 := {x Ū ψ(x) = 1}; (3) S 2 := {x / Ū ψ(x) = 1}. Note that S T, S 1. If S 2 = then l Ū; and Ū l satisfies the lemma. Assume that S 2 l. We claim that every line in Π(1, z, t) lies in the union of two of such subsets. Clearly, this holds for l. Let l Π(1, z, t) be a different line and put s := l l. If s Ū, then l(s, t) Ū, by construction, and all points s l S T are in ū and those with ψ(s) = 1 in Ū. In particular, l(s, t) S T S 1. If s l is in S 2, then l(s, x) is of type (F), ( F) or (C), and hence ψ takes at most two values on l(s, x), including ψ(s ) = 1. If s 2 l(s, x), ψ(s 2 ) = 1, x ū, ψ(x) 1, then s 2 S 2. Otherwise, if s 2 U, x ū, and then ψ is injective on l(s, x) = l(s 2, x), by the argument above. Hence s 2 S 2. Thus l(s 2, t ), with t l(s, t), is contained either in S 2 S T or S 1 S 2. Any y Π(1, z, t), with ψ(y) 1, is contained in ū. Indeed, consider l(y, y ), with ψ(y) ψ(y ), y l(t, s), ψ(y ) 1, and s y := l(y, y ) l(1, z). Then ψ(s y ) = 1, hence y /s y ū, and ψ is injective on l(y, y ). Since y ū, we find that y ū and s y S 1. Thus S T ū and any line l(y, s), with ψ(s) = 1, is either contained in S 2 S T or in S T S 1. This implies that any l(s, s 2 ), with s S 1, s 2 S 2, is contained in S 1 S 2. Note that none of the lines is contained in one of the subsets S T, S 1, S 2. By Proposition 5, the decomposition Π = S T S 1 S 2 is either (1) a cone over the decomposition of l(t, s) into the intersection with S T and S 1, and S 2 is just one point in l; (2) or is induced from a decomposition of P 2 (Z/p) over the residue of l, with S 1 equal to the preimage of a point, and hence S 2 l is projectively equivalent to Z (p).

HOMOMORPHISMS AND ALGEBRAIC DEPENDENCE 27 7. Proof of the main theorem We turn to the proof of Theorem 2, describing the homomorphisms ψ : P(K) P(L), preserving algebraic dependence. There are two possibilities: (V) ψ factors through a valuation, (P) ψ factors through a subfield, described in detail in the Introduction. We organize our proof as a case by case analysis, based on types of line, introduced at the beginning of Section 6. We consider two sets of cases as follows. Generic cases: ψ(u) (respectively, ψ(ū)), contains nonconstant algebraically independent elements, i.e., there exist y 1, y 2 ψ(u) (respectively, ψ(ū)) such that y 1 y 2. Degenerate cases: these sets do not contain algebraically independent elements. In our proof we need the following technical assumption: (AD) ψ(ū) does not contain nonconstant algebraically independent elements. This is satisfied when K has positive characteristic. In characteristic zero, this assumption allows us to avoid the case of geometric valuations which are induced from fields of positive characteristic. Lemma 32. Assume that ψ(u) contains nonconstant algebraically dependent elements and that P(K) contains lines of type (F) and possibly also (C). Then there exists a valuation ν of K such that o ν U and ψ((1 + m ν ) ) = 1. Proof. By Lemma 24, U K /k is a group, the induced quotient map K /k K /U is a nontrivial flag map, by the assumption on the existence of lines of type (F) in P(K) and by Proposition 5 und using Theorem 6 and Lemmas 7, 9. By Proposition 5, there is a map o µ K Γ µ, for some valuation µ, with the property that K K /U is a composition Let K r µ Γ µ K /U. Γ + µ := ν(o µ \ 0) Γ µ